apeNEXT: the future

 Schifano S. Fabio - INFN

Until APEmille the challenge was: build a processor performing as much as possible floating point operations

Today the main bottelneck is the node interconnection structure, so the focus is moved on:
how to build a network in order to fill as much as possible data to the processor?

If you build a good network the choise for the processor isn't so trivial as in the past and not unique.

Valid alternatives may be:

  1. custom processor: using standard cell technique design available frequence may be in the range of 200-300 Mhz --> 1,6 - 2,0 Gflops

  2. commercial processor: available frequence are around 1Ghz (Athlon, PIII) --> 2Gflops

We have chosen the first one and most of the old HW and SW will be reused and improved

Suivant